I am working on an application of LLVM where I would strongly prefer that getelementptr not be lowered into pointer arithmetic by any passes other than my own. I’m writing a ModulePass.
I am observing a situation where I compile a C++ file with no optimization and I get code as I would expect that uses getelementptr. However, if I enable optimization with -O3, some of my getelementptr instructions are lowered to ptrtoint/add-mulinttoptr sequences, but not others.
As far as I can tell, there is no advantage to what LLVM is doing here (such as re-using pointer arithmetic computations).
The ‘Often Misunderstood GEP page’ makes it sound like optimization always replaces pointer arithmetic with GEP instructions at the end, even if it might be an ‘uglygep’ (is this an instruction? it isn’t in the language reference) instead of a getelementptr. Is the problem that my pass is not at the end of the optimization passes? If this is the case, how can I more effectively guarantee this? Right now I’m adding my pass immediately before MachineModuleInfo (after the verifier pass, if it is enabled).
Another note, this only happens in a few programs I try to build, most of the time getelementptr is left unchanged through the compiler to my pass.
Is there a way to prevent LLVM from doing this? Btw, I’m using LLVM 3.0.